cheme-phi3-GGUF

ChemE-Phi3-GGUF

This repository contains GGUF (GPT-Generated Unified Format) quantized weights for ChemE-LLM, a domain-specific fine-tuned model based on microsoft/Phi-3-mini-4k-instruct. It is tailored specifically for chemical engineering simulation environments (DWSIM and MATLAB) and optimized for Retrieval-Augmented Generation (RAG) pipelines.

For the full open-source codebase, data curation pipelines, backend FastAPI server, and Next.js UI, visit our GitHub Repository.

Model Overview

• GitHub Codebase & Pipeline: bruhpika/ChemEng_finetuning-main
• Base Model: microsoft/Phi-3-mini-4k-instruct
• Domain Specialization: Chemical Engineering Simulations (DWSIM and MATLAB).
• Training Data: Supervised Fine-Tuning (SFT) on ~5,300 synthetic QA pairs generated from a dual-source knowledge base:
  1. Track A (Official Documentation): Verified technical manuals, documentation, academic papers, and HTML/PDF reference guides for DWSIM and MATLAB.
  2. Track B (Curated Media / YouTube Videos): Expert-curated instructional YouTube videos, visual tutorials, and procedural walkthroughs transcribed and structured into technical knowledge chunks.
• Knowledge Base (KB): The raw sources were deduplicated and chunked into 763 validated knowledge chunks (DWSIM: 296 chunks, MATLAB: 461+ chunks), which serve both as the foundation for training data synthesis and as the grounding database for RAG retrieval during live inference.
• Intended Use: Technical assistance, RAG-grounded QA, and step-by-step procedural guidance for chemical engineers.
• Context Window: 4,096 tokens

---

Quantization / Memory Ladder

Choose the GGUF file that best fits your hardware RAM/VRAM constraints:

---

Quickstart Guide

1. Running with llama-server / llama.cpp (Recommended)

You can launch an OpenAI-compatible API server using llama-server:

# Launch server on port 8081 with Q4_K_M weights
llama-server.exe -m cheme-phi3-q4_k_m.gguf -c 4096 --port 8081 -ngl 999

Query the server via curl:

curl http://127.0.0.1:8081/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "cheme-phi3",
    "messages": [
      {"role": "system", "content": "You are a chemical engineering assistant knowledgeable in DWSIM and MATLAB."},
      {"role": "user", "content": "How do I configure the parameters for a Flash Drum in DWSIM?"}
    ],
    "temperature": 0.2
  }'

2. Running with Ollama

Create a file named Modelfile in the same directory as the .gguf file:

FROM ./cheme-phi3-q4_k_m.gguf
PARAMETER temperature 0.2
PARAMETER num_ctx 4096
SYSTEM "You are an expert chemical engineering AI assistant trained in DWSIM and MATLAB workflows."

Create and run the model in Ollama:

ollama create cheme-phi3 -f Modelfile
ollama run cheme-phi3

---

Evaluation & Performance Note

When deployed alongside our domain-specific Vector Store (ChromaDB with 763 validated engineering documentation chunks), ChemE-Phi3 demonstrates high accuracy in determining thermodynamic properties, configuring unit operations, and generating clean simulation code while minimizing hallucinations.

License & Acknowledgements

• License: MIT License
• Lead Engineer: Harshith Bhardwaz Kenkari
• Acknowledgements: Built using QLoRA fine-tuning on microsoft/Phi-3-mini-4k-instruct and exported using llama.cpp.

Ollama Quick Start (Easiest Method)

For users who want to chat with the model immediately without setting up a Python virtual environment, you can use Ollama to pull and run the model directly from our Hugging Face repository in a single command. Depending on your hardware, you can choose from all available quantization tiers:

# 1. Run the recommended Q8_0 model (Best balance of speed/accuracy, ~4.06 GB)
ollama run hf.co/bruhpika/cheme-phi3-GGUF:Q8_0

# 2. Run the balanced Q5_K_M model (Excellent speed/accuracy, ~2.76 GB)
ollama run hf.co/bruhpika/cheme-phi3-GGUF:Q5_K_M

# 3. Run the ultra-compact Q4_K_M model (For older hardware/constrained devices, ~2.40 GB)
ollama run hf.co/bruhpika/cheme-phi3-GGUF:Q4_K_M

# 4. Run the unquantized F16 base model (Maximum fidelity, requires ≥12GB RAM, ~7.64 GB)
ollama run hf.co/bruhpika/cheme-phi3-GGUF:F16